Kronos: The First Open-Source Foundation Model Treating K-Line as a Language

Forum topic · ✨步子哥 · 2026-08-03

Summary

Kronos is the first open-source foundation model purpose-built for financial markets, framing K-line (candlestick) prediction as a language modeling problem. The article explains why generic time-series foundation models underperform on finance: high noise, non-stationarity, and heavy-tailed distributions. Kronos introduces a two-stage framework. Stage one uses a specialized hierarchical tokenizer that quantizes OHLCV data into multi-scale discrete tokens, preserving fine- and coarse-grained price movements that uniform binning would collapse. Stage two performs autoregressive Transformer pretraining over these tokens, analogous to GPT but with a financial vocabulary. Pretrained on K-line data from 45+ global exchanges, the released family spans Kronos-mini (4.1M params, 2k context), small, and base (all open source), with Kronos-large closed. The work was accepted to AAAI 2026 (arXiv:2508.02739). The article situates Kronos within the broader "X as a language" paradigm and notes engineering insights for tokenization, context length, and scaling.

Why Financial Time-Series Prediction Is Hard

General-purpose Time-Series Foundation Models (TSFMs) are trained on relatively clean signals such as electricity load or traffic flow. When applied to financial markets they underperform, because price data is high-noise, non-stationary, and heavy-tailed.

Kronos: A Two-Stage Framework

Kronos (AAAI 2026; arXiv:2508.02739; trending on GitHub) is described as the first open-source foundation model pretrained specifically for financial markets. Its core contribution is not a new architecture, but a new financial "language" via two stages:

1. Specialized hierarchical tokenizer for OHLCV — OHLCV (open/high/low/close/volume) is continuous and multi-dimensional. Rather than uniform binning, Kronos uses hierarchical quantization that captures price movements at multiple scales. Small moves (e.g., 1%) and large moves (e.g., 5%) therefore receive distinct token vocabularies, analogous to a technical analyst simultaneously reading daily, weekly, and monthly charts. 2. Autoregressive Transformer pretraining — Standard GPT-style autoregressive pretraining is applied over the discrete tokens. The "words" are K-line tokens instead of natural language.

Data and Model Family

Pretraining data covers K-line series from 45+ global exchanges, exposing the model to a broad "dialect" rather than a single market.

| Model | Parameters | Context length | Open source | |---|---|---|---| | Kronos-mini | 4.1M | 2048 | Yes | | Kronos-small | 24.7M | 512 | Yes | | Kronos-base | 102.3M | 512 | Yes | | Kronos-large | 499.2M | 512 | No |

Kronos-mini's compactness (4.1M parameters, 2k context) suggests that the financial K-line "language" has a smaller vocabulary and shorter dependencies than natural language.

Engineering Insight: "X as a Language"

Kronos exemplifies a wider pattern of converting continuous or raw signals into discrete tokens for Transformer pretraining:

NLP: text tokens
VLMs: image patches as tokens
Speech models: waveforms as tokens
Kronos: K-line as tokens

The article argues that the real strength of Transformers lies in any sequence that can be meaningfully tokenized, not only natural language.

Cross-Paper Resonance

The piece also ties Kronos to adjacent ideas: native multimodal scaling laws that emphasize dedicated tokenizers; position-encoding work (e.g., Mobius-style RoPE) that informs context-length choices; and an "edit the construction plan, not the blueprint" analogy, where Kronos's innovation sits in the data representation rather than the model architecture.

Resources

Code: https://github.com/shiyu-coder/Kronos
Paper: https://arxiv.org/abs/2508.02739
Live demo: https://shiyu-coder.github.io/Kronos-demo/
Models on Hugging Face: NeoQuasar/Kronos-{mini,small,base}